Large Scale GAN Training for High Fidelity Natural Image Synthesis

TL;DR

This paper presents BigGANs, a large-scale GAN model trained on ImageNet that achieves state-of-the-art high-fidelity, diverse image synthesis by applying orthogonal regularization and a truncation trick to control sample quality and variety.

Contribution

The paper introduces a new large-scale GAN training method with orthogonal regularization and a truncation trick, significantly improving image quality and diversity on ImageNet.

Findings

Achieved Inception Score of 166.5 on ImageNet at 128x128 resolution.

Reduced FID to 7.4, setting new state-of-the-art results.

Demonstrated the effectiveness of orthogonal regularization and truncation in large-scale GANs.

Abstract

Despite recent progress in generative image modeling, successfully generating high-resolution, diverse samples from complex datasets such as ImageNet remains an elusive goal. To this end, we train Generative Adversarial Networks at the largest scale yet attempted, and study the instabilities specific to such scale. We find that applying orthogonal regularization to the generator renders it amenable to a simple "truncation trick," allowing fine control over the trade-off between sample fidelity and variety by reducing the variance of the Generator's input. Our modifications lead to models which set the new state of the art in class-conditional image synthesis. When trained on ImageNet at 128x128 resolution, our models (BigGANs) achieve an Inception Score (IS) of 166.5 and Frechet Inception Distance (FID) of 7.4, improving over the previous best IS of 52.52 and FID of 18.6.